基于PSRGAN结合迁移学习的OCT视网膜图像超分辨率重建  

OCT retinal images super-resolution reconstruction based on PSRGAN and transfer learning

作　　者：陈明惠[1] 许诗怡 柯舒婷 邵怡[2] 吴玉全 CHEN Minghui;XU Shiyi;KE Shuting;SHAO Yi;WU Yuquan(Shanghai Engineering Research Center of Interventional Medical Device,University of Shanghai for Science and Technology,Shanghai 200093,China;Department of Urology,Shanghai General Hospital,Shanghai 200080,China)

机构地区：[1]上海理工大学上海介入医疗器械工程技术研究中心,上海200093 [2]上海市第一人民医院泌尿中心,上海200080

出　　处：《光学仪器》2024年第6期64-72,共9页Optical Instruments

基　　金：上海市科委产学研医项目(15DZ1940400)。

摘　　要：研究针对光学相干断层扫描(optical coherence tomography,OCT)图像采集中的斑点噪声和伪影问题,提出了渐进式超分辨率生成对抗网络(progressive super-resolution generative adversarial networks,PSRGAN)模型,并结合迁移学习方法来提高OCT视网膜图像的重建质量。PSRGAN模型以生成器和判别器组成的超分辨率生成对抗网络为框架,在判别器中加入改进的PECA模块,能够充分捕获多尺度特征图的空间信息,并实现图像跨维度通道特征的交互。实验结果显示,在峰值信噪比、结构相似性指数和边缘保留指数等指标中,该方法较PSRGAN–TL–X-ray网络分别提升了约2.19%、10.07%和4.64%,表明该方法相较于其他方法在图像质量和自动分割效果上有显著提升。To solve the problem of speck noise and artifacts in optical coherence tomography(OCT)image acquisition,a PSRGAN model is proposed and a transfer learning method is combined to improve the reconstruction quality of OCT retinal images.The PSRGAN model was based on the super-resolution generative adversary network(SRGAN)composed of generator and discriminator,and the improved PECA module is added to the discriminator,which can fully capture the spatial information of multi-scale feature maps and realize the cross-dimensional channel feature interaction of images.As for the peak signal-to-noise ratio(PSNR),structural similarity index(SSIM)and edge retention index(EPI),the proposed method had better results in comparison with the best performance PSRGAN–TL–X-ray network by 2.19%and 4.07%,10.64%,respectively.The results show that the proposed method significantly improves the image quality and automatic segmentation effect compared with other methods.

关 键 词：OCT图像 超分辨率 生成对抗网络 迁移学习 金字塔注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]